💻 computer science

TRACE: Thermal Recognition Attentive-Framework for CO2 Emissions from Livestock

The paper introduces TRACE, a novel thermal recognition framework that leverages mid-wave infrared video and specialized attention mechanisms to achieve state-of-the-art, non-invasive, and continuous quantification of CO2 emissions from free-roaming livestock through precise plume segmentation and flux classification.

Original authors: Taminul Islam, Abdellah Lakhssassi, Toqi Tahamid Sarker, Mohamed Embaby, Khaled R Ahmed, Amer AbuGhazaleh

Published 2026-04-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Taminul Islam, Abdellah Lakhssassi, Toqi Tahamid Sarker, Mohamed Embaby, Khaled R Ahmed, Amer AbuGhazaleh

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to listen to a single person whispering in a crowded, windy stadium. Now, imagine that person is a cow, the whisper is a cloud of invisible gas (CO2) coming out of its nose, and the stadium is a busy farm.

For a long time, scientists had two bad options to listen to this whisper:

The Cage: Put the cow in a tiny, air-tight box. This works, but it's stressful for the cow and you can only do it with one animal at a time.
The Bait: Trick the cow into coming to a special feeder to take a sample. This changes how the cow eats and acts, so the data isn't "real."

TRACE is a new, clever system that solves this problem without ever touching the cow. It uses special "heat-vision" cameras (like night-vision goggles that see heat) to watch cows roam freely. But here's the tricky part: CO2 gas is invisible to the naked eye, even to regular cameras.

The paper introduces TRACE (Thermal Recognition Attentive-Framework for CO2 Emissions from Livestock) as a super-smart AI detective that can "see" these invisible gas clouds and figure out how much the cow is "breathing out" just by watching a video.

Here is how TRACE works, broken down into simple analogies:

1. The Special Glasses (The Camera)

First, the system uses a special camera that sees a specific type of heat (infrared). Think of it like wearing glasses that turn invisible smoke into a glowing orange cloud. When a cow exhales, the camera sees a bright, swirling cloud of gas.

2. The "Gas-Sensing" Brain (TGAA)

The first job of TRACE is to draw a perfect outline around that glowing gas cloud in every single frame of the video.

The Problem: Gas clouds are weird. They are fuzzy, they change shape, and they look a lot like the background heat of the barn. Regular AI gets confused and either draws the outline too big or misses parts of it.
The TRACE Solution: Imagine the AI has a "gas-sense" superpower. It doesn't just look at the picture; it looks at the intensity of the gas at every single pixel.
The Analogy: Think of a spotlight. If you are looking for a specific person in a crowd, you usually scan everyone. TRACE is like a spotlight that automatically shines only on the person wearing a red hat (the gas). It ignores the rest of the crowd. This allows it to draw the outline of the gas cloud with extreme precision, even when the cloud is very faint or fuzzy.

3. The "Storyteller" (ATF)

The second job is to figure out what the cow is doing. Is it just resting? Is it digesting a big meal? Is it in a frenzy of fermentation?

The Problem: A single picture of a gas cloud doesn't tell the whole story. A gas cloud looks different when a cow is sleeping versus when it's eating. You need to see the movement over time to understand the story.
The TRACE Solution: This part of the AI is like a movie critic who watches the whole clip, not just one frame. It connects the dots between the gas clouds in frame 1, frame 2, and frame 3.
The Analogy: Imagine watching a movie of a balloon inflating. If you only look at one frame, you don't know if the balloon is being blown up or deflated. TRACE watches the whole sequence of the "breath cycle" (the inhale and exhale) to understand the cow's metabolic state. It's like listening to the rhythm of a song rather than just one note.

4. The Training Camp (The Curriculum)

Training an AI to do two hard things at once (drawing outlines and telling a story) is like trying to teach a student to play the piano and solve math problems simultaneously. They often get confused.

The TRACE Solution: The researchers used a "four-stage training camp."
1. Stage 1: Teach the AI to just draw the outlines (Segmentation).
2. Stage 2: Teach it to understand the story (Classification) using a "teacher" model to help it get started.
3. Stage 3 & 4: Let them work together, but carefully, so the math doesn't mess up the drawing.
The Analogy: It's like learning to ride a bike. First, you practice balancing (drawing the outline). Then, you practice pedaling (telling the story). Finally, you ride the whole bike, but you have training wheels that you slowly remove so you don't crash.

Why Does This Matter?

For the Farmer: It helps them know exactly how healthy each cow is. If a cow is producing too much gas, it might be sick or not eating the right food.
For the Planet: Cows produce a lot of greenhouse gases. By measuring exactly how much each cow emits, farmers can reduce waste and help the environment without hurting the animals.
The Result: TRACE is so good that it beats all other existing AI models, even though it is much smaller and lighter (like a sports car vs. a heavy truck). It can see the gas clearly, draw the perfect outline, and tell the farmer exactly what the cow is doing, all while the cow is just walking around freely.

In short: TRACE is a pair of magic glasses and a super-smart brain that lets us "see" invisible cow breath, measure it perfectly, and understand the cow's health, all without ever putting a collar on the animal.

1. Problem Statement

Quantifying exhaled $CO_2$ from free-roaming cattle is critical for precision nutrition, disease detection, and farm-scale carbon accounting. However, existing methods face significant limitations:

Respiration chambers and GreenFeed feeders require physical confinement or alter natural feeding behaviors, preventing scalable, continuous monitoring.
Portable breath samplers are sensitive to wind and distance, providing only point concentrations rather than volumetric flux.
Computer Vision Gap: While Mid-Wave Infrared (MWIR) thermal cameras can visualize $CO_2$ plumes (due to strong absorption at 4.2–4.4 µm) without chemical markers, no existing system can simultaneously perform per-frame plume segmentation and clip-level emission flux classification. Current models either ignore the temporal dynamics of the breath cycle or fail to handle the amorphous, low-contrast nature of thermal gas plumes.

2. Methodology: The TRACE Framework

TRACE is a unified framework designed to process MWIR thermal video clips ( $V = \{v_t\}_{t=1}^T$ ) and a co-registered per-pixel $CO_2$ intensity map ( $\Psi_t$ ). It outputs two tasks:

Per-frame binary plume segmentation ( $\hat{M}_t$ ).
Clip-level flux classification ( $\hat{y} \in \{\text{High-Flux, Control, Low-Flux}\}$ ).

The architecture consists of three core innovations:

A. Thermal Gas-Aware Attention (TGAA) Encoder

Instead of standard self-attention, TRACE replaces blocks in a Mix Vision Transformer (MiT-B0) with TGAA blocks that use the physical $CO_2$ intensity map ( $\Psi$ ) as a spatial supervisory signal.

Gas-Weighted Attention: Standard attention scores are modulated by per-patch gas intensity. High-intensity regions (dense plumes) receive higher attention weights, directing the model to focus on emission sources.
Spatial Dispersion Gate: A learned gate refines the aggregated values based on local gas concentration, helping the model distinguish between the plume and the complex thermal background.
Output: Produces multi-scale feature maps used for both segmentation and temporal fusion.

B. Attention-based Temporal Fusion (ATF) Module

To capture breath-cycle dynamics (essential for flux classification), the ATF module aggregates information from three parallel streams via cross-frame attention:

Stream A: Auxiliary mask features (from the segmentation decoder).
Stream B: Temporal frame descriptors derived from the TGAA encoder's final stage.
Stream C: A lightweight CNN representation.
The module uses cross-frame attention to align these streams, allowing the model to learn the temporal evolution of the plume without the overhead of processing every frame individually for classification.

C. Four-Stage Progressive Training Curriculum

To prevent gradient interference between the segmentation and classification objectives, TRACE employs a staged training strategy:

Stage S1(a): Warms up the segmentation decoder with the encoder frozen.
Stage S1(b): Unfreezes the TGAA encoder and trains it jointly with the decoder (segmentation only).
Stage S2: Uses a frozen VideoMAE-Small as a "teacher" to pre-align the ATF temporal stream via feature-matching loss. (VideoMAE is discarded after this stage).
Stage S3: End-to-end fine-tuning of the full pipeline with a joint loss function ( $L_{total} = \lambda_{seg} L_{seg} + \lambda_{cls} L_{cls}$ ), where segmentation is weighted higher to preserve localization precision.

3. Key Contributions

First Unified Framework: TRACE is the first system to jointly address per-frame gas plume segmentation and sequence-level flux classification from thermal video.
TGAA Mechanism: Introduces a novel attention mechanism that conditions self-attention on physical gas intensity, significantly improving boundary localization for amorphous objects.
Temporal Reasoning for Flux: Demonstrates that capturing breath-cycle dynamics via cross-frame attention is indispensable for distinguishing metabolic states (High vs. Low flux), a task single-frame models fail at.
Efficiency: Achieves state-of-the-art performance with only 4.1M parameters, outperforming models with up to 7x more parameters.

4. Experimental Results

The framework was evaluated on the CO2 Farm Thermal Gas Dataset (12 beef cattle, 5,040 frames, 432 clips) against 15 state-of-the-art baselines.

Segmentation Performance:
- mIoU: 0.9982 (Best in class).
- Boundary F1 (BF1): 0.9887 (Significantly outperforms gas specialists like Gasformer and CarboFormer).
- Centroid Localization Error (CLE): 0.021 pixels (10x better than specialized gas models).
- TRACE correctly delineates diffuse plume regions that baselines either over-segment, under-segment, or miss entirely.
Classification Performance:
- Accuracy: 0.827.
- Cohen's $\kappa$ : 0.741 (vs. $\leq 0.334$ for single-frame gas specialists).
- Gini Score: 0.882.
- The model successfully distinguishes between High-Flux (peak fermentation), Control, and Low-Flux (resting) states.
Ablation Studies:
- Removing TGAA caused BF1 to drop by ~49 points, proving gas-conditioned attention is critical for boundary precision.
- Removing ATF (temporal modeling) caused classification $\kappa$ to collapse by ~45 points, proving breath-cycle dynamics are essential for flux discrimination.
- Removing the S2 alignment stage reduced classification $\kappa$ by 5.9 points.

5. Significance and Impact

Non-Invasive Monitoring: TRACE enables continuous, per-animal $CO_2$ monitoring without physical contact, confinement, or chemical markers, making it viable for commercial-scale farms.
Metabolic Insight: By linking visual plume dynamics to metabolic states, the system offers a new tool for early disease detection and precision nutrition.
Carbon Accounting: Provides the necessary spatially resolved data for accurate farm-scale greenhouse gas accounting.
Efficiency: The framework demonstrates that domain-specific architectural choices (gas-aware attention) are more effective than simply scaling up model size or using generic video models.

Future Work: The authors note that generalization across different breeds, seasons, and wind conditions needs validation, and future iterations aim to include real-time edge inference and multi-animal tracking.